FairCauseSyn: Towards Causally Fair LLM-augmented Synthetic Data Generation.

Area of Science:

• Health Informatics
• Artificial Intelligence
• Data Science

Background:

• Synthetic data generation is crucial for health applications, requiring high-quality and fair data for equitable outcomes.
• Existing generative models (GANs, LLMs) often focus on counterfactual fairness, primarily in finance and legal fields, neglecting causal fairness in healthcare.
• Causal fairness offers a more robust evaluation framework by preserving data's causal structure, an aspect not addressed by current synthetic data methods in health.

Purpose of the Study:

• To develop the first LLM-augmented synthetic data generation method specifically designed to enhance causal fairness in real-world tabular health data.
• To address the gap in synthetic data generation methods that fail to incorporate causal fairness principles within health applications.

Main Methods:

• Developed a novel LLM-augmented approach for synthetic data generation.
• Applied the method to real-world tabular health data.
• Evaluated the generated data based on causal fairness metrics and its impact on bias reduction.

Main Results:

• The generated synthetic health data demonstrated less than 10% deviation from real data concerning causal fairness metrics.
• Training predictors on the causally fair synthetic data reduced bias related to sensitive attributes by 70% compared to using real data.
• The method successfully enhances causal fairness in synthetic health data.

Conclusions:

• This work introduces a pioneering LLM-augmented method for generating causally fair synthetic health data.
• The developed approach significantly improves fairness and reduces bias in sensitive attributes within health datasets.
• This advancement facilitates greater access to fair synthetic data, crucial for promoting equity in health research and healthcare delivery.